Woah, @sidsheth at @dMatrix_AI making the big moves You know the future of compute is heterogeneous when @JensenHuang starts putting down some chips Now we know why they integrated into nvidia's nvlink a few weeks back P.s you could have mentioned this at dinner last night buddy 😄

I actually need to raise 1b for spicy compute

SemiAnalysis numbers imply OpenAI is selling Cerebras-powered Ultrafast inference at ~$200M per megawatt per year. People are re-running the maths hoping for a typo. We're not surprised. A megawatt is a megawatt. What it earns depends on the silicon inside it. Here's how we think about it: Near term, revenue per megawatt wins. Power is the scarcest input in AI, and whoever monetizes each MW best gets the next MW. Long term, gross margin per megawatt wins. Prices compress, competitors arrive, and only the operators with real cost and pricing advantages keep the spread. Cerebras is built for both. Speed is the product. A frontier model at 750 tokens per second is something GPU clouds haven't publicly matched, and customers like Jane Street pay a premium for it. Premium pricing on a contracted, scarce resource is how you protect margin when everyone else is racing to the bottom. That's why we're extremely bullish on Cerebras, and why we're buying as many megawatts as we can. Premium inference unlocks the most delightful user experience in AI. No more waiting, waiting, waiting, staring at little thinking animations while the model works in the background. Answers that keep up with your thoughts. Our customers are even more excited than we are. That's General Compute.

Why I think owning decode silicon will make or break the inference giants like Baseten Fireworks and Together AI in 2027. My thoughts only // investment advice only to my VCs. The natural reaction to Kimi K3 pulling more or less level with the frontier labs has been to sell the frontier. And the reactionary logic makes sense: if open weights can get within a few points of Fable, the labs capture less value than everyone underwrote, and the winners end up being the companies distributing open source - Baseten, Fireworks, Together ai. But I've been thinking about this and I'm really not sure. The reason is in the license that Moonshot Kimi team launched rather than the benchmarks. K3 moving from open source to a constrained commercial contract is a bigger problem for inference providers who don't own their compute than it is for the frontier labs. Their whole business is arbitraging the gap between closed and open, and K3 just closed that gap from the wrong end. Hear me out.. Inference providers built the distribution layer for open source AI. That work is now being used against them. When Llama, DeepSeek, Qwen and Kimi shipped weights, they shipped a file. Not a product. Together, Fireworks and Baseten turned that file into something a developer could actually call - endpoints, uptime, quantization, tooling, billing. Without that channel, open weights stay a research artifact. And the channel got built for free, by companies whose entire pitch rested on one arbitrage: frontier-adjacent intelligence, at a fraction of closed-source price, with no vendor holding your neck. That pitch depended on two gaps. 1/ an intelligence gap they wanted to close, and 2/ a price gap they needed to stay open The intelligence gap closed (or is closing). K3 scores 57 on the Artificial Analysis Intelligence Index against Opus 5's 61. Fourth overall, open or closed. On cost per task it comes in below GPT-5.6 Sol and roughly half of Claude Opus 4.8. This was supposed to be the win condition. Then the price gap closed too. K2.6 was $0.95 in, $4 out. K3 is $3 and $15 :: a 3-4x generational increase that puts it in Claude Sonnet territory. Which tells you the open-weight discount was never really a property of open weights. It was customer acquisition, and that campaign is over? The license is what makes me think this is deliberate rather than incidental.. K3 didn't ship under MIT or Apache. It shipped under a bespoke Kimi K3 License. In it it says if you operate model-as-a-service and your group's aggregate revenue crosses $20M over any consecutive twelve months, you have to negotiate a separate agreement with Moonshot before any commercial use. Group revenue, not K3 revenue - so the trigger is met on day one by everyone who matters. Then look at who the license actually restricts. 1/ It exempts internal enterprise use. 2/ It explicitly excludes products that merely embed the model. Application builders and researchers are fine. 3/ The one business model gated behind a negotiated agreement is reselling raw inference. And Section 4 exempts "certified inference partners," which means Moonshot gets to decide who gets clean terms and who gets a phone call. I didn’t read that as a safety clause or an ideological one. It's more like a commercial instrument, pointed directly at the channel that built the category. Here's the part I think is under appreciated. This doesn't just compress margins. It takes price off the table as a customer acquisition tool entirely. If your input cost is set bilaterally, behind an NDA, by a supplier who also sells retail, you can't win a customer by being cheaper. That lever is just gone. Which leaves performance. And performance runs into a ceiling you can't software your way through. Speed, latency, sustained throughput, SLAs an agentic workload can actually plan against. That's where the real spread is right now. K3 benchmarks around 33 tokens per second across providers. Databricks reports 239 tokens/s on its own stack with a custom speculator. Identical weights, identical license, identical negotiated cost - call it a 14x difference in what the customer experiences. The obvious objection is that Databricks got there with software, not silicon, and that speculative decoding plus better kernels is a real moat. I'd argue the opposite. Software advantages in this market are measured in quarters, because the techniques publish and the talent moves. Silicon commitments are measured in years. Every software trick converges on the same ceiling eventually, and that ceiling is set by whatever the decode path physically runs on. The open-weight era made model choice free. It didn't make decode free. Once you can't compete on price, and software parity shows up on a two-quarter clock, the only durable question left is what silicon sits underneath your decode path. This is why we are betting the house on buying up all the decode silicon we can get our hands on. Starting with @SambaNovaAI SN50 machines. Obvious disclosure: this is the multi-hundred million dollar bet we are making.